Does NGD converge to minima that are 'more complex' i.e. have a higher RLCT?

{'model_type': 'dnn', 'experiment_type': 'standard', 'optimizer': 'both', 'hessian': True, 'hidden_nodes': [64, 64, 128, 128, 256, 256], 'hidden_layers': [1, 2, 1, 2, 1, 2], 'hidden_conv_layers': [1, 2, 3], 'cut_off_epochs': 20, 'num_epochs': 20, 'sgd_lr': 0.01, 'ngd_lr': 0.01, 'alpha': 0.01, 'eta': 0.9, 'epsilon': 1e-10, 'delta': 0.0005, 'momentum': 0.9, 'nesterov': True, 'seed': 1, 'batch_size': 128, 'num_workers': 6, 'dataset': 'mnist', 'num_hessian_batches': 1, 'sampler': 'sgld', 'num_chains': 1, 'num_draws': 1000, 'localization': 100.0, 'sampler_lr': 0.0001}